Take a sentence, translate it into English, send both versions to an AI. The French version will consume more tokens, so cost more, and take up more space in the model's working memory. For identical content. This is neither a bug nor deliberate discrimination: it's a direct consequence of how these systems split up text. Here's the explanation, and what you can do about it.
This is one of the most tangible blind spots of AI for French speakers. There's a lot of talk about cultural biases in models, less about this mechanical, quantifiable inequality that hits your bill and the limits of your conversations. Let's break it down.
Reminder: the token, the unit that matters
An AI doesn't read words, it reads tokens, those fragments of text that are its basic unit, as we explained in our dedicated article. Everything is billed per token, and the context window, that working memory we discussed in another article, is also measured in tokens.
The tool that does this splitting is called the tokenizer. And it's the one responsible for the inequality between languages.
Why French is at a disadvantage
A tokenizer isn't hand-programmed. It's learned from a corpus of texts, by identifying the most frequent character sequences and assigning them a unique token. The logic is one of compression: what comes up often gets a shortcut.
But training corpora are massively dominated by English. Direct consequence: common English words each have their own dedicated token, while French words are more often split into several pieces. The word "the" is one token. The word "anticonstitutionnellement" obviously takes several, but even mundane French words end up fragmented.
Three features of our language make things worse. Accents first: é, è, ç, ù are less frequent in corpora, so less well optimised, and an accented word can be split where its unaccented equivalent wouldn't be. Rich morphology next: our conjugations and agreements multiply the forms of a single word, which dilutes their individual frequency and reduces their chances of getting their own token. Finally, average length: French is structurally more verbose than English, with articles, prepositions and longer constructions to say the same thing.
In French, expect around one token per 0.7 words, whereas English tends to sit around one token per 0.75 words, with gaps that vary depending on the models and tokenizers. At the scale of a sentence, it's negligible. At the scale of an entire book fed to an AI, an agent running in a loop, or a company's monthly bill, the gap becomes a budget line. And you should know that a tokenizer change between two versions of the same model can shift consumption noticeably, which has surprised more than one developer when seeing their bill.
The three concrete consequences
You pay more. For identical content, a French text mechanically costs more than an English one, both in input and output. For personal use, it's painless. For a company processing millions of documents, it's a structural surcharge tied to the working language.
Your context window fills up faster. This is perhaps more annoying than the price. If a model accepts a million tokens, you'll fit fewer pages in French than in English. On a long conversation or a large document, you'll hit the limit sooner, and the model will lose sight of the start of the exchange before an English speaker would have that problem.
Quality can suffer at the margins. A word split into four fragments is a bit harder to process than a word represented by a single token. It's not dramatic on recent models, which are very solid in French, but it partly explains why performance often remains slightly better in English, especially on the most technical tasks.
What you can do
For everyday use, don't change anything. Writing in French, in the language you think in, produces better answers than forcing approximate English. The cost gap doesn't justify degrading your request.
For intensive or automated use, measure. If you're running an agent or processing large volumes, compare the real cost of your tasks in French and English on your specific cases. On certain technical processing, especially code where keywords are English anyway, switching system instructions to English can cut the bill without losing any quality. It's the same measurement reflex we recommended regarding the hidden cost of reasoning tokens.
Work on conciseness. A wordy prompt costs more than a precise one, and French naturally invites circumlocution. Getting straight to the point pays off doubly here.
What to remember
There's something revealing in this technical detail. These systems were built on a corpus dominated by English, and that dominance leaves traces right down to the most basic mechanics of word splitting. It's not malicious intent, it's a statistical imprint, and it has a measurable cost for everyone who thinks in another language.
The good news is that the gap is narrowing. Recent tokenizers are better trained on multilingual corpora, and European models are actively working on this efficiency. But as long as English remains the default language of training data, it will remain the cheapest language to speak with a machine. It's a small thing, and at the same time a useful reminder: technical neutrality doesn't exist, even in word splitting.